topic Re: extend based striping or kilobyte striping in Disk Enclosures

extend based striping or kilobyte striping

Michal Toth — Fri, 13 Jan 2006 10:39:41 GMT

Hi all,

simple question. What is better to employ for heavy loaded oracle server - distributed allocation policy (extend based striping) or to stripe on a kilobyte level (-i -I).

unfortunately, it is not possible to stripe on array level

thanks

Michal

Re: extend based striping or kilobyte striping

Patrick Wallek — Fri, 13 Jan 2006 11:05:17 GMT

That depends. Are the disks in this array raid protected in some fashion, RAID 1 or RAID 5?

If so, then I would do normal striping (-i and -I to lvcreate).

If these disks are NOT RAID protected then I would try distributed striping so you can mirror them.

Re: extend based striping or kilobyte striping

Michal Toth — Sat, 14 Jan 2006 21:14:08 GMT

Hi,

actually the underlying disk array uses several RAID5 groups to allocate space for luns. Bad thing is that it doesn't know how to strippe luns across these groups (IBM DS8300).

However I do not have to worry about mirroring.

I ask only from the performance point of view.

Re: extend based striping or kilobyte striping

Devender Khatana — Mon, 16 Jan 2006 04:44:13 GMT

Hi,

It will also depend upon the type of I/O your application performs at the LUNs. If it is sequential read majority, you should consider using extent based stripping. Anadditional advantage here is that you can load balance the controllers. This should be designed in such a way that half of the disks in a VG are accesses through one controller and other half through the other one. Also the allocation is done is such a way that first extent is on the disk accessed through first controller and the second is through the second controller and so on.

If the I/O size is too small you should consider KB level stripping. This also will depend across how much LUNs you are going to strip it across.

If the LUNs presented are RAID-5, these are allready stripped upto some extent and doing either on the desired stripping at OS level will help almost equal.

HTH,
Devender

Re: extend based striping or kilobyte striping

Michal Toth — Mon, 16 Jan 2006 06:01:31 GMT

well, the luns are protected by RAID-5. I think that the array presents 5 luns on each controller. So the option here is to either stripe across 10 pv or ditribute across 10 pv.

I know, that the application+oracle generates outstanding amount of i/o and it seems to be more random than sequential, because the system they use currently has low buffer cache hit ratio. I think that oracle block size is set to 8KB.

Re: extend based striping or kilobyte striping

Alzhy — Mon, 16 Jan 2006 12:47:32 GMT

If your IBM DS LUNS are already RAID Protected (RAID5 or something else) - striping those LUNS on the host level should give you added performance or might even be a best practice for your array like it is on ours (XP or Hitachi disks).

Striping on the host/volume manager (LVM or VxVM layer) will work best if you are sure that the LUNS were carved from different RAID groups inside your Disk Array and presented accross different disk chanels (Fibre or SCSI). I stripe no less than 4 ways but no more than 8 depending on the number of disk channels. As far as stripe size - I use 64KB for mixed use RDBMS storage (whetehr filesystem or raw). 32KB for OLTP and 128KB or more for DSS/WArehouse and other specialty applications:

LVM:

lvcreate -L 1024000 -i 8 -I 64 -n ora01 /dev/sapvg01

VxVM:

vxassist -g sapdg01 make ora01 100g layout=stripe ncols=8 stwidth=64

HTH.

Re: extend based striping or kilobyte striping

Michal Toth — Mon, 16 Jan 2006 16:28:01 GMT

Hi Nelson,

thanks for your lore, that's exactly something I was expecting to get. Anyway, the oracle uses 8KB db_block_size, vxfs reads in 32KB chunks, wouldn't it be better for stripe size to match one of these values?

thanks for insights

cheers

Mike

Re: extend based striping or kilobyte striping

Alzhy — Mon, 16 Jan 2006 16:32:58 GMT

Look for that VxFS tuning guide somewhere which explains certain tunables that may work wonders vis a vis your striping approach and your VxFS payload.

Generally I will say yes but you may get expert guidance out of that document.

Again - that is IF you are using cooked storage (meaning filesystem storage) for your Oracle storage.